ReWorld boosts autonomous driving WAMs with 24% accuracy gain
New framework improves autonomous driving video generation by 23.9% FVD...
ReWorld, the first representation learning framework specifically designed for autonomous-driving World Action Models (WAMs), directly optimizes intermediate representations across three dimensions: future-predictive supervision on the Video DiT, cross-modal alignment on the Action DiT, and hard-negative supervision for safety-critical boundaries. On nuScenes and NAVSIM, ReWorld improves fine‑tuned video generation by 23.9% in FVD (from 81.3 to 61.9), raises closed‑loop PDMS from 89.1 to 90.4 without any post‑training (e.g., RL or post‑processing), and accelerates from‑scratch convergence by approximately 2×.
- ReWorld improves autonomous driving WAMs' video generation by 23.9% FVD (81.3→61.9) on nuScenes/NAVSIM benchmarks
- Accelerates from-scratch training convergence by ~2x while boosting safety-critical planning accuracy (PDMS 89.1→90.4)
- First framework to directly optimize intermediate world representations across generation and planning modules
Why It Matters
ReWorld could reduce autonomous vehicle training costs by 50% while improving safety-critical decision accuracy—potentially accelerating deployment timelines.